[RISCV] Support non-power-of-2 types when expanding memcmp

llvm/lib/Target/RISCV/RISCVISelLowering.cpp

@@ -16191,10 +16191,6 @@ combineVectorSizedSetCCEquality(EVT VT, SDValue X, SDValue Y, ISD::CondCode CC,
     return SDValue();
 
   unsigned OpSize = OpVT.getSizeInBits();
-  // TODO: Support non-power-of-2 types.
-  if (!isPowerOf2_32(OpSize))
-    return SDValue();
-
   // The size should be larger than XLen and smaller than the maximum vector
   // size.
   if (OpSize <= Subtarget.getXLen() ||
@@ -16215,14 +16211,22 @@ combineVectorSizedSetCCEquality(EVT VT, SDValue X, SDValue Y, ISD::CondCode CC,
           Attribute::NoImplicitFloat))
     return SDValue();
 
+  assert(OpSize % 8 == 0 && "The size should be a multiple of 8");
   unsigned VecSize = OpSize / 8;
-  EVT VecVT = MVT::getVectorVT(MVT::i8, VecSize);
-  EVT CmpVT = MVT::getVectorVT(MVT::i1, VecSize);
+  EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i8, VecSize);
+  EVT CmpVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1, VecSize);
 
   SDValue VecX = DAG.getBitcast(VecVT, X);
   SDValue VecY = DAG.getBitcast(VecVT, Y);
-  SDValue Cmp = DAG.getSetCC(DL, CmpVT, VecX, VecY, ISD::SETNE);
-  return DAG.getSetCC(DL, VT, DAG.getNode(ISD::VECREDUCE_OR, DL, XLenVT, Cmp),
+  SDValue Mask = DAG.getAllOnesConstant(DL, CmpVT);
+  SDValue VL = DAG.getConstant(VecSize, DL, XLenVT);
+
+  SDValue Cmp = DAG.getNode(ISD::VP_SETCC, DL, CmpVT, VecX, VecY,
+                            DAG.getCondCode(ISD::SETNE), Mask, VL);
+  return DAG.getSetCC(DL, VT,
+                      DAG.getNode(ISD::VP_REDUCE_OR, DL, XLenVT,
+                                  DAG.getConstant(0, DL, XLenVT), Cmp, Mask,
+                                  VL),
                       DAG.getConstant(0, DL, XLenVT), CC);
 }
 

llvm/lib/Target/RISCV/RISCVTargetTransformInfo.cpp

@@ -2994,20 +2994,13 @@ RISCVTTIImpl::enableMemCmpExpansion(bool OptSize, bool IsZeroCmp) const {
   }
 
   if (IsZeroCmp && ST->hasVInstructions()) {
-    unsigned RealMinVLen = ST->getRealMinVLen();
-    // Support Fractional LMULs if the lengths are larger than XLen.
-    // TODO: Support non-power-of-2 types.
-    for (unsigned FLMUL = 8; FLMUL >= 2; FLMUL /= 2) {
-      unsigned Len = RealMinVLen / FLMUL;
-      if (Len > ST->getXLen())
-        Options.LoadSizes.insert(Options.LoadSizes.begin(), Len / 8);
-    }
-    for (unsigned LMUL = 1; LMUL <= ST->getMaxLMULForFixedLengthVectors();
-         LMUL *= 2) {
-      unsigned Len = RealMinVLen * LMUL;
-      if (Len > ST->getXLen())
-        Options.LoadSizes.insert(Options.LoadSizes.begin(), Len / 8);
-    }
+    unsigned VLenB = ST->getRealMinVLen() / 8;
+    // The minimum size should be `XLen / 8 + 1`, and the maxinum size should be
+    // `VLenB * MaxLMUL` so that it fits in a single register group.
+    unsigned MinSize = ST->getXLen() / 8 + 1;
+    unsigned MaxSize = VLenB * ST->getMaxLMULForFixedLengthVectors();
+    for (unsigned Size = MinSize; Size <= MaxSize; Size++)
+      Options.LoadSizes.insert(Options.LoadSizes.begin(), Size);
   }
   return Options;
 }